This is an audit you can finish in an afternoon. Pick one Linux host, ideally a mid-importance one, and work down the list. Each item states what to check and what a bad answer looks like. By closing time you should hold a short, written, risk-ranked list of fixes. The structure follows the scope that the Ops Atlas assigns to SysOps in its reference page on the discipline: services, privileges, remote access, logs, timers, packages, filesystems, time, the host firewall, backups and triage.
Before you start
- Open a second session on the host and leave it open. If a check requires a reload, you want a way back in.
- Create a scratch file for findings. Record the command, the result and your verdict.
- Do not change anything during the first pass. Look, note, then fix in a second pass.
Section A: access and privilege
- List accounts that can log in. Any shared or unnamed account is a finding.
- Run the dump of effective remote-login settings from the daemon’s test mode and compare it with what you think you configured. In this daemon the first value usually wins, so a stray early line can override a careful later one.
- Confirm that password login and direct administrator login are off, if your policy says so.
- Inspect the sudo rules directory. Files must be mode 0440, and any file with a dot in its name is skipped silently. Run the editor’s check mode, which also verifies ownership.
- Where the host is reached through a cloud session service, confirm that the default shared user has been replaced by per-person run-as settings.
Section B: scheduled work
- List every cron entry and every active timer. Note who owns each one.
- Search cron lines for a bare percent sign; it will be treated as a newline.
- For each important job, find the last successful run. A job with no recorded success is a finding, even if it has no recorded failure.
- For timers, check for persistence, a randomised delay and a failure hook. Missing ones are improvements, not emergencies.
Section C: logs and time
- Check the journal’s storage setting. If it is unset, the effective behaviour depends on whether a persistent directory exists.
- Check the size cap. By default the journal may take up to a tenth of the filesystem and leaves a larger share free.
- Review log rotation. Copy-and-truncate rotation can lose a very small slice of lines.
- Run the time service’s status. Is it synchronised? Is the offset small? What is the policy for large corrections?
Section D: packages and restarts
- List pending updates and the oldest one outstanding.
- Check whether a restart is owed: the Red Hat family helper returns exit code one when it is, and Ubuntu leaves a marker file.
- Know your undo. dnf can reverse a transaction from its history; apt cannot, and you would pin an older version.
- Validate the mount table and test mounting everything before the next reboot.
- Look up the end-of-support date of the operating system release and write it down.
Section E: protection layers
- Is the host firewall on, and does the rule set match the services the host should expose?
- Is SELinux or AppArmor in enforcing mode? Any host-wide permissive setting is a finding. Per-domain or per-profile exceptions are acceptable if they are documented.
- For firewall edits, confirm there is a procedure that reverts automatically unless you confirm the change.
Section F: capacity and limits
- For each filesystem, compute days of runway from the growth rate, including the share reserved for the administrator. As a worked case, a 500 GiB volume with 410 GiB used and 9 GiB daily growth has about seven days, not ten.
- Check the open-file limit of the busiest services. The default soft limit is 1,024.
- If a process has been tuned so the kernel will never kill it, note it: the Atlas warns that this only shifts the kill elsewhere.
Section G: recovery
- Find the last backup and the last test restore. No test means the backup is unproven.
- Confirm that snapshots are not being counted as backups.
- Check that you can reach the machine without its network, through a console path, and that any disk-encryption recovery keys are stored somewhere other than on the machine.
Section H: triage readiness
- Is there a script, stored with the runbook, that prints load, memory, disk, network and top-process data in a fixed order? If not, write one.
- Is historical performance data being collected, so a bad hour can be compared with a good one?
- Do alerts fire on a few signals tied to user pain, rather than on dozens of raw counters?
- Would the on-call person know not to restart first? Add the sentence to the runbook if not.
Readings that trigger false alarms
An audit also needs a short list of things that look wrong and are not. A load figure higher than the number of processors may reflect tasks waiting on storage, not saturated processors. A small amount of free memory is normal because the system caches aggressively; judge by the available-memory figure. A disk showing full busy time on flash storage or a striped array may still have headroom. Write these false alarms next to the finding they could be mistaken for, so the next auditor does not spend an hour on them.
How to write up a finding
Keep each finding to four lines: what you saw, why it matters, the smallest safe fix and who owns it. Avoid language such as “should be improved”. Say “remove the shared login by the end of the month” and name the person. Findings written this way get closed; findings written as essays get filed.
Scoring and follow-up
Tally findings by section and rank them: access problems first, because they widen every other risk; unproven recovery second; drift and housekeeping after. The Atlas suggests a handful of measures for tracking progress over time: baseline conformance, monitoring coverage, scheduled-job success, disk runway, reboot debt, toil share and end-of-life exposure. Pick three and record them monthly.
If your team also runs automated incident tooling, two related reads on this site cover the response side of the picture: how smarter incident handling speeds up resolution and a practical roadmap for operations teams adopting automation. Use them to decide where host findings should feed alerts.
Repeat the audit on a different host next week. By the fourth round you will have a feel for the typical failure patterns, and a standard template for the rest of the estate.

